Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
INTRODUCTION GRAPH APPROACH EXPERIMENTS CONCLUSION
COMPUTATION OF RENAMEABLEHORN BACKDOORS
Stephan Kottler, Michael Kaufmann andCarsten Sinz
University of Tuebingen, Germany
15th May 2008 @ SAT’08 Guangzhou
1
INTRODUCTION GRAPH APPROACH EXPERIMENTS CONCLUSION
OUTLINE
1 INTRODUCTION2 GRAPH APPROACH
Theoretical BackgroundGreedy HeuristicApproximation
3 EXPERIMENTSComparing ResultsSimplification of Dependency Graphs
4 CONCLUSION
2
INTRODUCTION GRAPH APPROACH EXPERIMENTS CONCLUSION
INTRODUCTION
Backdoor Set: particular subset of variablesWilliams, Gomes and Selman 2003
Focus on variables of a strong backdoor issufficient to decide satisfiability
Example of real-world instance: 6,700 vars &440,000 clauses→ backdoor with 12 variables
Random instances have much larger backdoors(Interian)
Finding a minimum backdoor is hard (Szeider)
3
INTRODUCTION GRAPH APPROACH EXPERIMENTS CONCLUSION
DELETION BACKDOORS
Introduced by Nishimura, Ragde and Szeider
Defined with respect to a base class C
C recognizable and solvable in poly. timee.g. base classes Horn, 2-SAT, Renameable Horn
B ⊆ V is a deletion backdoor if F −B belongs to CF −B: remove from F all occurrences of variables (pos./neg.) in B
A deletion backdoor is a strong backdoor if C isclause-induced
4
INTRODUCTION GRAPH APPROACH EXPERIMENTS CONCLUSION
RECENT WORK
Paris et al.: RHorn-Backdoors for ZChaff
Try to rename variables to increase the number ofHorn clauses (WalkSat)Greedily choose variables for the backdoor
Dilkina, Gomes and Sabharwal:Computed optimal Backdoors for different baseclassesMinimum Renameable Horn Backdoors≤ Minimum Renameable Horn Deletion Backdoors
5
INTRODUCTION GRAPH APPROACH EXPERIMENTS CONCLUSION
RHORN AS A GRAPH PROBLEM
For formula F we create a graph G = (V ,E):V: Each variable xi entails two vertices:
k0i represents that xi has to be renamedk1i represents that xi must not be renamed
. . . to make F HornE: Represent the implications of renaming
or not renaming variables (according to clauses)
EXAMPLE : (xi ∨xj ∨ . . .)If xj is renamed than xi has also to be renamedIf xi is not renamed than xj must not be renamed
6
INTRODUCTION GRAPH APPROACH EXPERIMENTS CONCLUSION
CONFLICT LOOPS ANDCONFLICT SETSA variable xi has a conflict loop if there is a path from k0ito k1i and vice versa. The set of variables involved in aconflict loop is a conflict set.
A formula F is Renameable Horn iff there existsno variable that has a conflict loop in the graph.
Lewis proved: For any formula F there is a2-SAT fromula that is satisfiable iff F is RenameableHornThe Dependency Graph is the Implication Graph(Aspvall et al.) of Lewis’ 2-SAT-instance
7
INTRODUCTION GRAPH APPROACH EXPERIMENTS CONCLUSION
REDUCTION LEMMASIMPLIFICATION OF THE GRAPH
If variable xi does not have a conflict loop thenneither vertex k0i nor vertex k
1i can be involved
in a conflict loop of any other variable.
8
INTRODUCTION GRAPH APPROACH EXPERIMENTS CONCLUSION
RHORN DELETION BACKDOORS. . .AS A GRAPH PROBLEM
Goal: We have to get rid of all conflict loops!
Delete variables to remove conflict loops.
Why not using strongly connectedcomponents?⇒ At the beginning there is only one SCCTwo approaches to destroy conflict loops
9
INTRODUCTION GRAPH APPROACH EXPERIMENTS CONCLUSION
A GREEDY HEURISTICTO DESTROY ALL CONFLICT LOOPS
Function greedyRHornBackdoor (F )G = (V ,E)← Dependency Graph of FS← computeConflictSets (G,V )B← /0 (start with an empty backdoor)while S 6= /0 do
xi ← choose variable according to heuristicB← B ∪{xi}delete vertices k0i ,k
1i and incident edges
U← variables, whose conflict loops were destroyedS← S ∪ computeConflictSets (G,U)Apply reduction rules according to Reduction Lemma
return B
10
INTRODUCTION GRAPH APPROACH EXPERIMENTS CONCLUSION
APPROXIMATIONOF OPTIMAL RENAMEABLE HORN DELETION BACKDOORS
Function approxRHornBackdoor (F )G = (V ,E)← Dependency Graph of F ; B← /0while G contains at least one conflict loop do
C← vertices of one (preferably small) conflict loopB← B ∪{var(k) : k ∈ C}Hide all vertices related to vars in B (and incident edges)Apply reduction rules according to Reduction Lemma
forall x ∈ B doReinsert vertices (and edges) related to xif G contains no conflict loop then B← B \{x}else Undo reinsertion of vertices and edges related to x
return B
Inspired by an algorithm for the FEEDBACK ARC SET problem(Demetrescu & Finocchi)
11
INTRODUCTION GRAPH APPROACH EXPERIMENTS CONCLUSION
COMPARISON. . .TO LOCAL SEARCH APPROACHES
Number of WalkSat Dependency GraphInstance Vars Cls (Paris) heuristic 2-phase
apex7_w5 1500 11695 740 663 3.93s 627 1.60sc499_w5 2070 22470 885 837 5.33s 818 2.35sdp10s10 8372 23004 2635 1449 26.90s 1498 2.08slisa20_2 1201 6563 820 774 0.87s 799 0.22srand_net40 2000 5921 811 665 2.79s 692 0.34svda_w9 6498 130997 4809 4488 6m 4293 5mvmpc_21 441 45339 439 437 1.50s 424 10.13svmpc_25 625 76755 603 610 5.48s 605 46.17s
!
Graph Approach is independent of the number ofrenamings!
12
INTRODUCTION GRAPH APPROACH EXPERIMENTS CONCLUSION
SIMPLIFICATION OF THE GRAPHSFOR EASY INSTANCES
13
INTRODUCTION GRAPH APPROACH EXPERIMENTS CONCLUSION
SIMPLIFICATION OF THE GRAPHSFOR HARD INSTANCES
14
INTRODUCTION GRAPH APPROACH EXPERIMENTS CONCLUSION
PRACTICAL RELEVANCERESULTS FOR INSTANCES OFSAT COMPETITION 2007
Number of Heuristic ApproximationInstance Vars Cls BD % Time BD % Time
AProVE07-06 46335 632886 4485 9% 28.90s 4376 9% 12.24seq.a.braun.12 1694 5726 639 37% 36.59s 634 37% 1.07seq.a.braun.13 2010 6802 765 38% 45.17s 755 37% 1.86sdspam_vc1080 118298 375379 32018 27% 289m 40220 33% 78mmizh-md5-47-3 65604 273522 15077 22% 25m 16687 25% 1m
QG6-ukn2726 2123 9177 710 33% 15.47s 491 23% 30.18sbqwh.40.520 2211 14710 1431 64% 8.28s 1458 65% 0.64scontest02-26 744 2464 376 50% 0.13s 351 47% 0.08sgensys-ukn002 2129 8961 702 32% 18.26s 483 22% 29.01s
unif-k3-r4.25 450 1912 238 52% 0.76s 243 54% 0.28sunif-k7-r89 75 6675 74 98% 0.15s 74 98% 0.19sunif2p-p0.7 3500 9344 1065 30% 1m 1116 31% 1munif2p-p0.9 1170 4234 525 44% 3.87s 552 47% 3.32s
!
Industrial / Crafted / Random Instances
15
INTRODUCTION GRAPH APPROACH EXPERIMENTS CONCLUSION
CONCLUSION
What we did . . .
Two Approaches to compute Renameable Horn(Deletion) Backdoors
Realistic for small but hard instances!
Open Problems
How can Backdoors be used for the solvingprocess?
Fast Computation of Non-Deletion Backdoors
16